Deep learning models are notoriously overconfident when encountering data they have never seen before. To fix this, a popular paradigm emerged: Outlier Exposure (OE). The intuition behind OE is straightforward—if you want a model to reject anomalies at test time, expose it to a massive auxiliary dataset of anomalies during training so it learns to draw a boundary around the known, in-distribution (ID) data.
However, this approach suffers from a fundamental scalability bottleneck. The space of "everything else" is infinitely large. Sourcing diverse, representative outlier datasets is computationally expensive, legally or ethically challenging, and often impossible in specialized domains like medical imaging or malware detection. As noted by Jiefeng Chen et al. in Robust Out-of-distribution Detection in Neural Networks, relying on large auxiliary datasets creates a severe dependency that limits deployment in data-scarce environments.
To break this dependency, a growing body of research focuses on eliminating or drastically reducing the need for external auxiliary data. Instead of collecting the world's anomalies, these methods extract, synthesize, or geometrically infer the boundaries of the unknown from the ground up.
Finding Outliers Within: Intrinsic Proxies
If we cannot collect external outliers, can we find them hidden inside our existing data?
Consider a standard image classification dataset. Every image consists of a target object (the in-distribution concept) and a background. While the object belongs to a known class, the background contains arbitrary, non-target elements—sky, roads, or random textures.
Sen Pei leverages this insight in Image Background Serves as Good Proxy for Out-of-distribution Data. By exploiting the local properties of convolutional layers, the proposed framework isolates non-object background patches during training and treats them as a "free" proxy for out-of-distribution (OOD) data. Because backgrounds naturally mirror the statistical complexity of the real world, the model learns to draw tight boundaries around target objects without ever seeing an external anomaly.
A similar philosophy applies to few-shot scenarios where data is incredibly scarce. Atsuyuki Miyai et al. introduce LoCoOp in LoCoOp: Few-Shot Out-of-Distribution Detection via Prompt Learning, using CLIP's local features to identify background noise and explicitly push it away from target text prompts.
[Standard Image] ---> Isolate Object (In-Distribution)
---> Isolate Background (Intrinsic OOD Proxy) ---> Push Away
This reuse of intrinsic data also extends to the structure of training loops. In data-starved regimes, Nikhil Mehta et al. propose using "out-of-episode" samples—data points belonging to other classes within a meta-learning pipeline—as immediate outliers to sharpen decision boundaries without external data collection.
Synthesizing and Amplifying Sparse Anomalies
What if we only have access to a tiny handful of outlier examples? Standard outlier exposure fails here because the model quickly overfits to those specific examples, failing to generalize to novel anomalies.
To solve this, researchers focus on informative extrapolation. Jianing Zhu et al. developed Diversified Outlier Exposure (DivOE) in Diversified Outlier Exposure for Out-of-Distribution Detection via Informative Extrapolation. DivOE takes a limited set of auxiliary outliers and dynamically "stretches" them toward the in-distribution decision boundary during optimization. By synthesizing more challenging, near-boundary outliers, the model learns a robust boundary from minimal data. A similar hybrid strategy is pursued by Junwei He et al. for graph neural networks, interpolating between external graphs and synthesized internal outliers to map out fuzzy decision boundaries.
Taking this a step further, some methods eliminate real images entirely. The authors of Textual Training for the Hassle-Free Removal of Unwanted Visual Data: Case Studies on OOD and Hateful Image Detection demonstrate that vision-language models can be trained to detect OOD visual data using zero real images, relying instead on synthetic text descriptions to define the anomalous space.
Alternatively, Jinggang Chen et al. introduce a data-free approach (C2IR) that inverts a pre-trained network to synthesize class-specific "impressions." This recovers internal statistics to flag anomalies without needing the original training set or external outliers.
Geometry and Activation Shaping Over Data Collection
Instead of modifying the training data, another philosophy looks at how in-distribution and out-of-distribution data flow through a network's internal geometry. OOD data often triggers distinct, anomalous activation patterns that can be caught post-hoc without any extra training data.
For instance, Andrija Djurisic et al. introduce Activation SHaping (ASH) in Extremely Simple Activation Shaping for Out-of-Distribution Detection. They observed that OOD data often propagates low-magnitude, noisy activations through the network because it lacks the specific semantic features the network was optimized to detect. By simply pruning up to 90% of these weak activations at inference time, they clean up the signal and dramatically improve OOD detection without needing global thresholds or training data access.
OOD Input ---> [Layer Activations] ---> Prune Weak Signals (90%) ---> Clean OOD Score
Other researchers focus on the geometric spaces where features reside. Eduardo Dadalto Câmara Gomes et al. apply information geometry, replacing standard Euclidean metrics with the Fisher-Rao distance on probability manifolds to naturally separate distributions.
To capture complex semantic structures, Jeng-Lin Li et al. and Ming-Ching Chang et al. map features onto joint hyperspherical and hyperbolic spaces. Hyperbolic geometry is uniquely suited for hierarchical data, allowing the network to naturally push unfamiliar, out-of-distribution concepts into the expansive outer edges of the manifold, achieving the performance of massive outlier exposure models with a fraction of the data.
Drawing Boundaries from Inliers Alone
Ultimately, the cleanest way to avoid outlier exposure dependency is to build detectors that look exclusively at the structure of the inliers.
Masoud Taghikhah et al. demonstrate this with QuantOD in Quantile-Based Maximum Likelihood Training for Outlier Detection. Instead of trying to model what an outlier looks like, they train a normalizing flow using a quantile-based maximum likelihood objective that focuses heavily on the "hardest" inliers at the very edge of the known distribution. By precisely mapping the outer crust of the in-distribution data, the model naturally rejects anything falling outside it.
Similarly, Bertrand Charpentier proposes the Natural Posterior Network: Deep Bayesian Predictive Uncertainty for Exponential Family Distributions, which models uncertainty analytically for exponential family distributions, bypassing the need for synthetic or real OOD exposure during training.
As Philipp Liznerski demonstrates in Exposing Outlier Exposure: What Can Be Learned From Few, One, and Zero Outlier Images, massive outlier corpora are often an unnecessary crutch. By mastering the density, geometry, and internal representations of the known, models can robustly identify the unknown simply by knowing themselves.